Consensus disagreement analysis

Consensus Disagreement Analysis

When skills/consensus-synthesis detects a non-unanimous verdict for a journey, this skill resolves the disagreement by applying skills/sequential-analysis to each diverging criterion. It is the arbitration layer between vote tabulation and final verdict emission.

The authoritative contract for when this skill runs, and what resolutions it may emit, is rules/consensus-engine.md §Disagreement Protocol and §Synthesis States. In any conflict between this SKILL.md and the rule, the rule wins.

This skill is strictly read-only against validator subdirectories. It NEVER writes to validator-N/. It writes only to e2e-evidence/consensus/disagreement-analysis/.

When This Skill Runs

Triggered when consensus-synthesis computes a journey state of anything other than UNANIMOUS_PASS or UNANIMOUS_FAIL:

• MAJORITY_PASS → resolve whether the minority FAIL was validator error, missing evidence, or a real defect the majority missed.
• MAJORITY_FAIL → resolve whether the minority PASS was premature or evidence-weak, or whether the majority FAIL is wrong.
• SPLIT → attempt resolution; if unresolvable, return DISAGREEMENT_UNRESOLVED (LOW confidence).

Do NOT invoke this skill when the journey is UNANIMOUS — there is no disagreement to resolve, and forcing the protocol wastes cycles and risks fabricating a dissent that does not exist.

Protocol

Step 1: IDENTIFY    Step 2: LOAD         Step 3: INVOKE           Step 4: APPLY
Which criterion     Both validators'     sequential-analysis      Phases 1–4
did validators      evidence files for   with FAIL as symptom,    (Symptom → Hypothesize
rate differently?   the diverging        PASS evidence as         → Investigate → Conclude)
                    criterion            hypothesis to test

                                        Step 5: EMIT RESOLUTION
                                        MINORITY_CORRECT  → flip majority
                                        MAJORITY_CORRECT  → keep majority, record dissent
                                        UNRESOLVABLE      → DISAGREEMENT_UNRESOLVED

Step 1: Identify the Diverging Criterion

Read the per-criterion tabulation table emitted by consensus-synthesis. For a journey with K PASS criteria, find the rows where at least one validator disagreed with the others. A journey-level MAJORITY or SPLIT always has at least one diverging criterion — if it does not, the tabulation is malformed and must be regenerated before this skill proceeds.

#	Criterion	V1	V2	V3	Agreement
1	Login submits valid creds	PASS	PASS	PASS	UNANIMOUS (skip)
2	Error shown on bad pwd	PASS	PASS	FAIL	DIVERGING
3	Session persists refresh	PASS	FAIL	FAIL	DIVERGING

Emit e2e-evidence/consensus/disagreement-analysis/step-01-diverging-criteria.md listing each diverging criterion, the vote tuple, and which validators are in the majority vs the minority for that criterion.

Criterion-level, not journey-level. Two validators can both emit journey PASS while disagreeing on criterion #3 internally, if one counted criterion #3 as a non-blocker. Always operate on the criterion grain — that is what the per-criterion tabulation is for.

Step 2: Load Both Validators' Evidence (Read-Only)

For each diverging criterion, enumerate the evidence files cited by each side from their respective validator-N/ subdirectories:

e2e-evidence/consensus/validator-1/step-04-error-banner.png
e2e-evidence/consensus/validator-1/step-05-network-response.json
e2e-evidence/consensus/validator-3/step-04-error-banner.png
e2e-evidence/consensus/validator-3/step-05-console-error.txt

Open, read, and summarize each file. Do not modify any file inside validator-N/. Writes into another validator's evidence directory invalidate the independence guarantee of the whole consensus run — see rules/consensus-engine.md §File Ownership.

If a cited evidence file is missing or zero-byte, that is a first-class signal: the validator's verdict for this criterion is validator error (rules/consensus-engine.md §Disagreement Protocol case d). Record it and skip to Step 5 with a RESOLUTION of MAJORITY_CORRECT (minority discarded as validator error) — do NOT fabricate evidence to fill the gap.

Emit e2e-evidence/consensus/disagreement-analysis/step-02-evidence-inventory.md with one subsection per diverging criterion listing:

• Majority evidence files (path + one-line observation)
• Minority evidence files (path + one-line observation)
• Obvious gaps (missing files, zero-byte files, unreachable citations)

Step 3: Invoke skills/sequential-analysis

For each diverging criterion, call skills/sequential-analysis with this framing:

• Symptom (Phase 1 input): the FAIL verdict — "Validator-X reports FAIL for criterion {C} in journey {J}. Expected: {PASS criterion text}. Actual: {FAIL validator's cited observation}."
• Hypothesis to test (Phase 2 seed): the PASS verdict's evidence — "The PASS validators cited {files} which purport to show criterion {C} passing. Is this evidence sufficient, contradicted, or stale?"

Because the analysis is now comparative (PASS evidence vs FAIL evidence), the hypothesis set must include at minimum:

#	Hypothesis	How to confirm
H1	FAIL is correct; PASS evidence is stale/misread	Re-read PASS evidence against criterion wording; look for version/timestamp drift
H2	PASS is correct; FAIL captured a pre-fix state or flake	Check FAIL timestamps vs PASS timestamps; look for "works on retry" signal
H3	Criterion is under-specified; both readings are valid	Re-read criterion against both sides' citations; check for ambiguity
H4	FAIL validator made a reachability/tooling error	Check FAIL evidence for empty files, broken selectors, wrong URLs
H5	PASS validator made an evidence-quality error	Check PASS evidence for missing observations, bare status codes, 0-byte files

Additional hypotheses from the categories in skills/sequential-analysis (Data / Timing / Environment / Code / Integration / Infrastructure / Validation setup) are required when they apply.

Save the sequential-analysis inputs and thought log to e2e-evidence/consensus/disagreement-analysis/step-03-sequential-analysis-inputs.md.

Step 4: Apply Sequential-Analysis Phases 1–4

Execute the four phases defined in skills/sequential-analysis/SKILL.md within this consensus context. The phases map one-to-one:

Sequential-Analysis Phase	Consensus Context
Phase 1: SYMPTOM	Diverging criterion: exact PASS/FAIL wording + cited evidence on each side
Phase 2: HYPOTHESIZE	The comparative hypothesis table above, plus any category-specific additions
Phase 3: INVESTIGATE	Read both sides' evidence; confirm or rule out each hypothesis using only the already-captured files (no re-running of validators from inside this skill)
Phase 4: CONCLUDE	Root cause: which side's evidence is stronger, or whether the criterion is under-specified

Write the analysis report to e2e-evidence/consensus/disagreement-analysis/step-04-analysis.md in the format skills/sequential-analysis/SKILL.md specifies (Root Cause → Evidence Chain → Hypotheses Evaluated → Recommended Fix / Prevention).

Re-runs are the coordinator's job, not this skill's. If Phase 3 investigation concludes that a validator must re-capture evidence (case a or b in rules/consensus-engine.md §Disagreement Protocol), record the request and return control to the coordinator. This skill must not spawn a validator.

Step 5: Emit a Resolution

Each diverging criterion resolves to exactly one of three states. The criterion's resolution rolls up into the journey's final verdict per the rules in consensus-synthesis.

Resolution	Condition	Journey Effect
MINORITY_CORRECT	Phase 4 concludes the minority validator's evidence is stronger — the majority missed a real defect, misread the criterion, or captured stale state.	Flip the journey verdict to match the minority. The final verdict reflects the dissent; confidence stays at the agreement-ratio tier (MEDIUM at best, because the ratio was < 1.0 to begin with).
MAJORITY_CORRECT	Phase 4 concludes the majority's evidence is stronger — the minority validator erred (missing evidence, wrong selector, stale snapshot, misread criterion).	Keep the majority verdict. Record the dissenting validator's evidence and the analysis in the report so the dissent is never silently dropped.
UNRESOLVABLE	Phase 4 cannot decide — criterion is under-specified, evidence is genuinely contradictory, or the real system is non-deterministic and no re-run is feasible in this run.	Journey resolves to DISAGREEMENT_UNRESOLVED (LOW confidence). Escalate to human with the full analysis. Do NOT downgrade to MAJORITY to avoid the escalation.

Emit e2e-evidence/consensus/disagreement-analysis/step-05-resolution.md with one row per diverging criterion:

#	Criterion	Majority	Minority	Resolution	Rationale
2	Error shown on bad pwd	PASS (V1,V2)	FAIL (V3)	MAJORITY_CORRECT	V3 captured 0-byte screenshot; cited evidence unreadable.
3	Session persists refresh	PASS (V1)	FAIL (V2,V3)	MINORITY_CORRECT	V1's screenshot is from pre-refresh state; V2+V3 captured post-refresh cookie clear.

The resolution set is returned to consensus-synthesis, which re-synthesizes the journey with the post-analysis verdicts and finalizes the report.

No fourth state. The three resolutions above are exhaustive. Inventing a "MOSTLY_CORRECT" or "INCONCLUSIVE" bucket to dodge the UNRESOLVABLE escalation is an anti-pattern; see consensus-synthesis on the prohibition of inconclusive.

Evidence Path

All artifacts from this skill live under:

e2e-evidence/consensus/disagreement-analysis/
  step-01-diverging-criteria.md
  step-02-evidence-inventory.md
  step-03-sequential-analysis-inputs.md
  step-04-analysis.md
  step-05-resolution.md

Additional investigation artifacts (e.g., diffs between validators' screenshots, terminal transcripts) go in the same directory with step-NN-* prefixes so ordering is preserved.

This directory is owned by this skill. consensus-synthesis reads from it but does not write here; validators never touch it.

Integration

Skill / File	Relationship
rules/consensus-engine.md	Authoritative contract. §Disagreement Protocol specifies when this skill runs and what resolutions are permitted. In conflict, rule wins.
skills/consensus-synthesis	Invokes this skill whenever a journey's state is not UNANIMOUS. Consumes this skill's resolution set to produce the final synthesized verdict.
skills/sequential-analysis	Called in Step 3–4. This skill is a specialization of sequential-analysis: the FAIL verdict is the symptom, the PASS evidence is the seed hypothesis, and the resolution maps back to the consensus state machine.
agents/consensus-synthesizer	Agent that drives consensus-synthesis and therefore this skill. Owns the final write to e2e-evidence/consensus/report.md.
agents/verdict-writer	This skill inherits verdict-writer discipline: every resolution cites evidence; no dissent is silently dropped; no fabricated verdicts.

Iron Rules (Carry-Over)

1. Read-only against validator subdirectories. NEVER modify another validator's evidence — doing so invalidates the run.
2. No fabricated verdicts. If Phase 4 cannot decide, emit UNRESOLVABLE and escalate. Do not invent a tiebreaker.
3. No silent dissent. Even MAJORITY_CORRECT resolutions record the dissenter's evidence and the analysis reasoning.
4. No re-synthesis inside this skill. Return resolutions; let consensus-synthesis re-tabulate.
5. No validator spawning inside this skill. If a re-run is required, return that as an output and let the coordinator handle it.
6. No confidence upgrades. A resolved MAJORITY is still MEDIUM; flipping MINORITY_CORRECT does not become HIGH. Agreement ratio gates confidence.
7. Criterion-grain analysis only. Operate on the per-criterion tabulation; journey-level disagreement is always decomposable into one or more diverging criteria.